EPL 데이터 분석 및 클러스터링

데이터 : 2020 epl game data

데이터 전처리

import pandas as pd
import numpy as np

import plotly.express as px
import plotly.graph_objects as go
import matplotlib.pyplot as plt
import seaborn as sns

from sklearn.cluster import KMeans
from sklearn.preprocessing import MinMaxScaler
from sklearn.decomposition import PCA

import warnings

pd.set_option('display.max_rows',100)
warnings.filterwarnings('ignore')

data = pd.read_csv('C:/Users/hyoseok/Desktop/data_science practice/epl_data_analysis/epl2020.csv')
data.head()

	Unnamed: 0	h_a	xG	xGA	npxG	npxGA	deep	deep_allowed	scored	missed	...	AF.x	AC.x	AY.x	AR.x	B365H.x	B365D.x	B365A.x	HtrgPerc	AtrgPerc	matchDay
0	1	h	2.234560	0.842407	2.234560	0.842407	11	5	4	1	...	9	2	2	0	1.14	10.0	19.00	0.466667	0.416667	Fri
1	2	a	0.842407	2.234560	0.842407	2.234560	5	11	1	4	...	9	2	2	0	1.14	10.0	19.00	0.466667	0.416667	Fri
2	3	a	3.183770	1.200300	2.422640	1.200300	9	1	5	0	...	13	1	2	0	12.00	6.5	1.22	0.600000	0.642857	Sat
3	4	h	1.200300	3.183770	1.200300	2.422640	1	9	0	5	...	13	1	2	0	12.00	6.5	1.22	0.600000	0.642857	Sat
4	5	h	1.340990	1.598640	1.340990	1.598640	4	6	1	1	...	19	4	1	0	1.95	3.6	3.60	0.230769	0.375000	Sat

5 rows × 45 columns

data.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 576 entries, 0 to 575
Data columns (total 45 columns):
 #   Column        Non-Null Count  Dtype  
---  ------        --------------  -----  
 0   Unnamed: 0    576 non-null    int64  
 1   h_a           576 non-null    object 
 2   xG            576 non-null    float64
 3   xGA           576 non-null    float64
 4   npxG          576 non-null    float64
 5   npxGA         576 non-null    float64
 6   deep          576 non-null    int64  
 7   deep_allowed  576 non-null    int64  
 8   scored        576 non-null    int64  
 9   missed        576 non-null    int64  
 10  xpts          576 non-null    float64
 11  result        576 non-null    object 
 12  date          576 non-null    object 
 13  wins          576 non-null    int64  
 14  draws         576 non-null    int64  
 15  loses         576 non-null    int64  
 16  pts           576 non-null    int64  
 17  npxGD         576 non-null    float64
 18  teamId        576 non-null    object 
 19  ppda_cal      576 non-null    float64
 20  allowed_ppda  576 non-null    float64
 21  matchtime     576 non-null    int64  
 22  tot_points    576 non-null    int64  
 23  round         576 non-null    int64  
 24  tot_goal      576 non-null    int64  
 25  tot_con       576 non-null    int64  
 26  Referee.x     576 non-null    object 
 27  HS.x          576 non-null    int64  
 28  HST.x         576 non-null    int64  
 29  HF.x          576 non-null    int64  
 30  HC.x          576 non-null    int64  
 31  HY.x          576 non-null    int64  
 32  HR.x          576 non-null    int64  
 33  AS.x          576 non-null    int64  
 34  AST.x         576 non-null    int64  
 35  AF.x          576 non-null    int64  
 36  AC.x          576 non-null    int64  
 37  AY.x          576 non-null    int64  
 38  AR.x          576 non-null    int64  
 39  B365H.x       576 non-null    float64
 40  B365D.x       576 non-null    float64
 41  B365A.x       576 non-null    float64
 42  HtrgPerc      576 non-null    float64
 43  AtrgPerc      576 non-null    float64
 44  matchDay      576 non-null    object 
dtypes: float64(13), int64(26), object(6)
memory usage: 202.6+ KB

데이터 용어

1. xG : Expected Goals, 골 기댓값, 기대 득점, 슛이 득점으로 연결될 확률
2. xGA : Expected Goals Against, 실점 기댓값, 기대 실점
3. NPxG : 페널티 킥을 제외한 xG, 모든 공격수를 공평하게 평가할 수 있다.
4. NPxGA : 페널티 킬을 제외한 xGA
5. deep : Number of plays in opponent final third, 파이널 서드 지역에서의 기회
6. deep_allowed : Number of plays allowed in final third
7. scored : goals scored 득점
8. missed : goals conceded 실점
9. xPts : 기대 승점
10. ppda_cal : 압박 플레이를 한 정도
11. allowed_ppda : 압박 플레이를 당한 정도
12. tot_goal : 지금까지 팀이 득점한 총 골수
13. tot_con : 지금까지 팀이 실점한 총 골수
14. HS.x : Home team shots
15. HST.x : Home shots on target
16. HF.x Home fouls
17. HC.x :Home corners
18. HY.x : Home yellow card
19. HR.x : Home red card
20. AS.x : Away shots
21. AST.x : Away shots on target
22. AF.x :Away fouls
23. AC.x : Away corners
24. AY.x : Away yellow card
25. AR.x : Away red card
26. HtrgPerc : Shot on target/total shots - Home
27. AtrgPerc : Shot on target/total shots - Away

# Home 경기만 추출
home = data[data['h_a']=='h']

home.head()

	Unnamed: 0	h_a	xG	xGA	npxG	npxGA	deep	deep_allowed	scored	missed	...	AF.x	AC.x	AY.x	AR.x	B365H.x	B365D.x	B365A.x	HtrgPerc	AtrgPerc	matchDay
0	1	h	2.234560	0.842407	2.234560	0.842407	11	5	4	1	...	9	2	2	0	1.14	10.00	19.00	0.466667	0.416667	Fri
3	4	h	1.200300	3.183770	1.200300	2.422640	1	9	0	5	...	13	1	2	0	12.00	6.50	1.22	0.600000	0.642857	Sat
4	5	h	1.340990	1.598640	1.340990	1.598640	4	6	1	1	...	19	4	1	0	1.95	3.60	3.60	0.230769	0.375000	Sat
6	7	h	0.909241	1.087520	0.909241	1.087520	0	9	3	0	...	12	7	0	0	2.62	3.20	2.75	0.400000	0.272727	Sat
7	8	h	0.871590	1.224600	0.871590	1.224600	5	5	0	0	...	14	2	1	1	3.00	3.25	2.37	0.333333	0.300000	Sat

5 rows × 45 columns

# 문자열 컬럼 삭제 
home = home.drop(columns=['Unnamed: 0','h_a','result','date','teamId','Referee.x','matchDay'])

home.head()

	xG	xGA	npxG	npxGA	deep	deep_allowed	scored	missed	xpts	wins	...	AST.x	AF.x	AC.x	AY.x	AR.x	B365H.x	B365D.x	B365A.x	HtrgPerc	AtrgPerc
0	2.234560	0.842407	2.234560	0.842407	11	5	4	1	2.3863	1	...	5	9	2	2	0	1.14	10.00	19.00	0.466667	0.416667
3	1.200300	3.183770	1.200300	2.422640	1	9	0	5	0.2522	0	...	9	13	1	2	0	12.00	6.50	1.22	0.600000	0.642857
4	1.340990	1.598640	1.340990	1.598640	4	6	1	1	1.0172	0	...	3	19	4	1	0	1.95	3.60	3.60	0.230769	0.375000
6	0.909241	1.087520	0.909241	1.087520	0	9	3	0	1.1422	1	...	3	12	7	0	0	2.62	3.20	2.75	0.400000	0.272727
7	0.871590	1.224600	0.871590	1.224600	5	5	0	0	1.0512	0	...	3	14	2	1	1	3.00	3.25	2.37	0.333333	0.300000

5 rows × 38 columns

표준화

scaler = MinMaxScaler()
scaler.fit(home)
scaled = scaler.transform(home)
print(scaled)

[[0.32468965 0.17681474 0.36768374 ... 0.71842317 0.46666667 0.41666667]
 [0.16580485 0.70199763 0.18776005 ... 0.00321802 0.6        0.64285714]
 [0.18741789 0.3464427  0.21223501 ... 0.09895414 0.23076923 0.375     ]
 ...
 [0.29308506 0.08788059 0.33189419 ... 0.1230893  0.64705882 0.33333333]
 [0.24854551 0.11918188 0.2814569  ... 0.01850362 0.5        0.28571429]
 [0.3520082  0.13869184 0.26620376 ... 0.27594529 0.46666667 0.25      ]]

PCA(Principal Comeponent Analysis, 주성분 분석)

차원 축소 기법 중 하나로 Column의 갯수가 많을 때 Dimension을 낮추는 역할을 한다.

#PCA 설정
pca = PCA(n_components = 4)
pca.fit(scaled)

# PCA 적용
pca_result = pca.transform(scaled)
pca_result = pd.DataFrame(data=pca_result, columns = ['PCA1','PCA2','PCA3','PCA4'])

pca_value = pd.DataFrame(data = pca.components_.T, index=home.columns)
print(pca_value)

                     0         1         2         3
xG           -0.094734  0.015808 -0.118161  0.036680
xGA           0.140346 -0.131545  0.158305 -0.067885
npxG         -0.096503  0.017722 -0.126467  0.034830
npxGA         0.137182 -0.132086  0.170624 -0.062214
deep         -0.061460  0.001557 -0.153594  0.059538
deep_allowed  0.081760 -0.085197  0.230949 -0.027762
scored       -0.118833 -0.030096  0.010956  0.014804
missed        0.106785 -0.057480 -0.019667  0.025020
xpts         -0.252673  0.132415 -0.292995  0.089978
wins         -0.537585 -0.341442  0.181727  0.001402
draws         0.106972  0.753849  0.172270 -0.127088
loses         0.430613 -0.412407 -0.353998  0.125686
pts          -0.501928 -0.090159  0.239151 -0.040961
npxGD        -0.123258  0.070268 -0.157451  0.050147
ppda_cal      0.063279 -0.054920  0.163821 -0.052282
allowed_ppda -0.040542  0.017905 -0.152153  0.053959
matchtime     0.002313 -0.038000 -0.066778 -0.101766
tot_points   -0.060472 -0.031667 -0.145421 -0.301273
round        -0.023017 -0.045469 -0.171484 -0.649686
tot_goal     -0.063739 -0.043758 -0.207971 -0.362749
tot_con       0.032944 -0.049566 -0.083156 -0.488086
HS.x         -0.090384  0.057853 -0.239817  0.096210
HST.x        -0.124876  0.002992 -0.116020  0.031758
HF.x         -0.011263  0.022994  0.001764  0.000996
HC.x         -0.045859  0.063416 -0.201973  0.088770
HY.x          0.033301  0.044271  0.022442 -0.017600
HR.x          0.004588 -0.051744  0.066248  0.019078
AS.x          0.120768 -0.112266  0.300448 -0.077558
AST.x         0.107344 -0.103698  0.124777 -0.012166
AF.x          0.014959  0.005116  0.038348 -0.030747
AC.x          0.037019 -0.052745  0.165289 -0.082769
AY.x          0.009486  0.048210 -0.025238 -0.016834
AR.x         -0.016473  0.053703  0.045608  0.005274
B365H.x       0.083985 -0.079951  0.145808 -0.021150
B365D.x      -0.032822 -0.043475 -0.088837  0.033000
B365A.x      -0.085006  0.008618 -0.179389  0.058222
HtrgPerc     -0.053094 -0.040464  0.069991 -0.026693
AtrgPerc      0.045000 -0.039110 -0.084746  0.046488

• 득점율 : NPxG , scored, deep
• 실점율 : NPxGA. deep_allowed, missed
• 많은 득점 기회를 얻는 팀 : ppda_cal, NPxG, deep
• 많은 실점 기회를 갖는 팀 : allowed_ppda, NPxGA, deep_allowed

home1 = home[['npxG','scored','deep','npxGA','deep_allowed','missed']]
home1.head()

	npxG	scored	deep	npxGA	deep_allowed	missed
0	2.234560	4	11	0.842407	5	1
3	1.200300	0	1	2.422640	9	5
4	1.340990	1	4	1.598640	6	1
6	0.909241	3	0	1.087520	9	0
7	0.871590	0	5	1.224600	5	0

# 표준화
scaler = MinMaxScaler()
scaler.fit(home1)
scaled = scaler.transform(home1)

• Scree plot을 그려 몇개의 클러스터로 나눌지 확인
• Elbow method : 군집 수를 늘려가면서, 각 그룹내의 요소들이 얼마나 가까이 뭉처있는지 파악하는 방법

Scree plot을 기준으로 급격한 기울기로 꺾이는 지점을 elbow point라고 하며 이 때의 군집 수를 사용

def elbow(X):
  sse = []

  for i in range(1, 11):
    km = KMeans(n_clusters=i, init='k-means++', random_state=0)
    km.fit(X)
    sse.append(km.inertia_)

  plt.plot(range(1,11), sse, marker='o')
  plt.xlabel('Number of Clusters')
  plt.ylabel('Within Group Sum of Squares')
  plt.show()

elbow(scaled)

km = KMeans(n_clusters=4).fit(scaled)
print(km.cluster_centers_)

[[0.16742899 0.16111111 0.16953405 0.40391392 0.34656085 0.17037037]
 [0.19595621 0.1        0.21039427 0.1869128  0.18095238 0.08888889]
 [0.16943311 0.13333333 0.17204301 0.64665691 0.63492063 0.32962963]
 [0.39884896 0.34294872 0.35525227 0.1769826  0.20695971 0.07264957]]

predict = pd.DataFrame(km.predict(scaled))

predict.columns = ['predict']

scaled = pd.DataFrame(data=scaled)

result = pd.concat([scaled, predict], axis=1)

result.columns = ['npxG','scored','deep','npxGA','deep_allowed','missed','Cluster']

result['teamId'] = data['teamId']

result

	npxG	scored	deep	npxGA	deep_allowed	missed	Cluster	teamId
0	0.367684	0.500	0.354839	0.198122	0.238095	0.111111	3	Liverpool
1	0.187760	0.000	0.032258	0.595293	0.428571	0.555556	2	Norwich
2	0.212235	0.125	0.129032	0.388191	0.285714	0.111111	0	Man City
3	0.137126	0.375	0.000000	0.259728	0.428571	0.000000	0	West Ham
4	0.130576	0.000	0.161290	0.294181	0.238095	0.000000	1	Bournemouth
...	...	...	...	...	...	...	...	...
283	0.135209	0.000	0.064516	0.064026	0.190476	0.000000	1	Man City
284	0.558776	0.125	0.193548	0.080775	0.190476	0.111111	3	Aston Villa
285	0.331894	0.500	0.290323	0.098471	0.142857	0.000000	3	Chelsea
286	0.281457	0.250	0.096774	0.133544	0.095238	0.000000	1	Leicester
287	0.266204	0.500	0.258065	0.155405	0.095238	0.000000	3	Man Utd

288 rows × 8 columns

cluster_result = km.cluster_centers_
cluster_result = pd.DataFrame(data=cluster_result,
                              columns=['npxG','scored','deep','npxGA','deep_allowed','missed'])
cluster_result.head()

	npxG	scored	deep	npxGA	deep_allowed	missed
0	0.167429	0.161111	0.169534	0.403914	0.346561	0.170370
1	0.195956	0.100000	0.210394	0.186913	0.180952	0.088889
2	0.169433	0.133333	0.172043	0.646657	0.634921	0.329630
3	0.398849	0.342949	0.355252	0.176983	0.206960	0.072650

• 해당 시즌의 강팀인 'Liverpool'과 'Man City'
• 해당 시즌의 약팀인 'Norwitch'와 'Bournemouth'를 예시로 사용하였다.

# Liverpool 팀 정보 추출
Liv = result[result['teamId']=='Liverpool']
Liv = Liv.mean()
Liv = pd.DataFrame(data=[Liv[:-1]], columns=['npxG','scored','deep','npxGA','deep_allowed','missed'])

Liv.head()

	npxG	scored	deep	npxGA	deep_allowed	missed
0	0.27114	0.276786	0.251152	0.272995	0.272109	0.119048

# Man City 팀 정보 추출
Mc = result[result['teamId']=='Man City']
Mc = Mc.mean()
Mc = pd.DataFrame(data=[Mc[:-1]], columns=['npxG','scored','deep','npxGA','deep_allowed','missed'])

Mc.head()

	npxG	scored	deep	npxGA	deep_allowed	missed
0	0.211718	0.183333	0.23871	0.293671	0.304762	0.133333

# Norwich 팀 정보 추출
Nor = result[result['teamId']=='Norwich']
Nor = Nor.mean()
Nor = pd.DataFrame(data=[Nor[:-1]], columns=['npxG','scored','deep','npxGA','deep_allowed','missed'])

Nor.head()

	npxG	scored	deep	npxGA	deep_allowed	missed
0	0.240372	0.178571	0.193548	0.317782	0.251701	0.134921

# Bournemouth 팀 정보 추출
Bou = result[result['teamId']=='Bournemouth']
Bou = Bou.mean()
Bou = pd.DataFrame(data=[Bou[:-1]], columns=['npxG','scored','deep','npxGA','deep_allowed','missed'])

Bou.head()

	npxG	scored	deep	npxGA	deep_allowed	missed
0	0.330171	0.2	0.225806	0.218424	0.203175	0.081481

def Visualization(cluster, team, team_name):
  categories = cluster.columns

  fig = go.Figure()

  fig.add_trace(go.Scatterpolar(
      r = cluster.values[0],
      theta = categories,
      fill='toself',
      name='중하위권'
  ))

  fig.add_trace(go.Scatterpolar(
      r = cluster.values[1],
      theta = categories,
      fill='toself',
      name='중상위권'
  ))

  fig.add_trace(go.Scatterpolar(
      r = cluster.values[2],
      theta = categories,
      fill='toself',
      name='하위권'
  ))
  fig.add_trace(go.Scatterpolar(
      r = cluster.values[3],
      theta = categories,
      fill='toself',
      name='상위권'
  ))


  fig.add_trace(go.Scatterpolar(
      r = team.values[0],
      theta = categories,
      fill='toself',
      name=team_name
  ))

  fig.update_layout(
    polar=dict(
      radialaxis=dict(
        visible=True,
        range=[0, 0.7]
      )),
  )

  fig.show()

Visualization(cluster_result, Liv, 'Liverpool')

Visualization(cluster_result, Mc, 'Man City')

Visualization(cluster_result, Nor, 'Norwitch')

Visualization(cluster_result, Bou, 'Bournemouth')

Result

• 클러스터링이 제대로 되지 않은 것 같다.(약팀의 npxG값이 낮고, 강팀의 npxG값이 높아야하지만 그렇지가 않다.)
• 아마 표준화 단계에서 값이 많이 바뀌어서 그런 것 같다.
• 데이터 값에 대한 의미를 좀 더 파악하고 축구 점유율 과 같은 데이터가 추가되면 팀의 유형과 클러스터링을 하기 더 좋아질 것 같다.

참고 자료 및 출처

참고 자료 : https://western-sky.tistory.com/42?category=847897 데이터 : https://www.kaggle.com/datasets/idoyo92/epl-stats-20192020